fix: support non-unique handles for timers

[guybedford]

This was necessary to get one of the WPT cases to pass for Fastly's implementation.

Our implementation of checking task deadlines for timer tasks uses a redundant singular invalid handle (-2) to indicate a timer handle for a non-existent subscription. If we gave every timer a unique handle we would be using up the real handle ID space.

But the cancel_task currently assumes handle uniqueness when finding the handle.

This supports non-unique handles for tasks by taking the task itself to cancel_task instead of its handle to look it up from, storing timers as a map from timer ids to task pointers.

I can confirm this is working correctly in the WPT test case for Fastly.

[tschneidereit]

If I understand the problem correctly, all timers in Fastly's implementation have PollableHandle == -2 because of how timeouts are translated, right?

If so, that makes sense to me, yeah.

The tricky bit, and why I didn't immediately implement things this way, is that the TimerTasks in the map need to be properly rooted—see the comment in-line.

[guybedford]

If I understand the problem correctly, all timers in Fastly's implementation have PollableHandle == -2 because of how timeouts are translated, right?

Yes exactly.

I've implemented a very naive singleton and trace model here as a start. Further feedback very much welcome.

[tschneidereit]

Looks good! I left a bunch of suggestions which should be applied (and then a little bit of other tweaks as follow-ups to those, because I didn't add suggestions for all the places where s/timer_ids_/timers_/g needs to happen), but with those in, this'll be ready to go!

[tschneidereit]

Oh huh, seems like that doesn't work with raw pointers, sorry!

The suggested edits do make it work, because tracing is implemented for js::UniquePtr.